What happened
Anthropic launched a HackerOne bug bounty program specifically for cyber jailbreaks of Fable 5's classifiers, part of its redeployment agreement with the US government. Vetted researchers can attempt to bypass the cybersecurity classifiers, and successful submissions receive financial rewards and are treated as responsible disclosures.
The scope is narrowly defined to cyber jailbreaks rather than general model-safety misuse. Anthropic also published a draft jailbreak severity framework, co-developed with AWS, Microsoft and Google, that scores attacks by accessibility and harm.
Why it matters
Inviting outside researchers to probe a frontier model's safety classifiers is a shift toward the kind of adversarial testing long used in traditional security. A shared severity framework across major cloud providers could standardize how the industry rates these attacks.
The narrow scope keeps the program focused on the cyber risks regulators care about most.
MintedBrain take
Structured bounties and a common severity scale are healthy signs that model safety is being treated like real security work. For teams building on Fable 5, the resulting disclosures are worth following as a read on where the classifiers are weakest.
Discussion
Sign in to comment. Your account must be at least 1 day old.